What data for this category looks like
Typically, production ledger data for semiconductor wafer fabrication comes from MES systems, equipment FDC data, defect detection and metrology tools, wafer maps, process recipe libraries, and experimental data. Update cycles proceed alongside individual wafer batch production, real-time parameter collection, and daily aggregation. Daily documents include process flow cards, batch operation records, equipment recipes, SPC control charts, exception handling sheets, defect maps, and yield reports. Fields include wafer batch numbers, nanoscale process parameters, equipment operating temperatures, defect position coordinates, and yield values. Units involve nanometers, degrees Celsius, and piece counts.
What constraints these characteristics impose on the document parsing and chunking link
Ultra-long process chain documents must be chunked without interrupting process logic, avoiding splitting a single process step across multiple pages. Structured data such as nanoscale parameters and defect coordinates requires that chunking retains complete parameter groups, preventing separation of parameters from their corresponding coordinates. Real-time updated equipment and batch data requires the parsing process to adapt to fast processing needs, avoiding timeout interruptions. Differences in structure across multiple document types (such as the row-based structure of Excel ledgers and the paginated structure of PDF process cards) require flexible adjustment of chunking rules to adapt to the layout characteristics of different documents.
How to set up the configuration
This configuration applies to FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale for this value |
|---|---|---|
chunkSize | 800–1200 characters | Wafer fabrication documents include nanoscale process parameters and long process chains, requiring single chunks to contain complete parameter groups or single process step logic |
chunkSplitter | Custom delimiters (based on line breaks, process numbers, batch numbers) | Different document types (Excel ledgers, PDF process cards) have large structural differences; custom delimiters avoid cross-row splitting or process step breaks |
PARSE_FILE_TIMEOUT_SECONDS | 600–900 seconds | Sufficient time must be reserved for parsing 10M-level large files to avoid timeout interruptions |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Compatible with complete ledger packages for single wafer batches |
chunkSettingMode | Manual mode | Chunking rules need to be adjusted for different document types (SPC charts, defect maps); manual mode supports flexible configuration |
Three common mistakes
- Parsing 10M-level Word process ledgers takes excessive time, and the interface shows stalled parsing progress. The cause is failure to adjust the
PARSE_FILE_TIMEOUT_SECONDSparameter; the default timeout threshold is insufficient for large-volume wafer ledger files. - Excel-imported equipment recipe ledgers are automatically split into multiple rows within the same chunk, with each equipment parameter not treated as a separate chunk. The cause is failure to configure
chunkSplitterto use line breaks or batch numbers as delimiters; the default chunking logic does not adapt to table row structures. - Chunks for yield reports of specific batches cannot be retrieved, but re-importing the same file allows normal retrieval. The cause is a cache exception triggered during the first parsing, and the complete document parsing and chunking process was not completed.
How to confirm the configuration is correct
- Upload a single 10M-level Word process ledger, check whether parsing duration meets expectations, and confirm that
PARSE_FILE_TIMEOUT_SECONDSis set sufficiently to cover the parsing duration. - Import an Excel equipment recipe file, check whether each equipment parameter is treated as a separate chunk, and confirm that
chunkSplitteris configured correctly. - Randomly select 3 documents of different types (process flow cards, exception handling sheets, yield reports) to verify that chunked content retains complete process or parameter logic.
- Search the imported knowledge base to confirm that key parameters such as wafer batch numbers and etching temperatures can be normally retrieved, verifying that chunking configuration does not lose core information.
The values provided are common starting points and should be measured against the reader's own samples.